6

Q: Let $m, n$ be positive intergers. Let $A$ and $B$ real $n\times n$ matrices. Assume that $B$ is symmetric and positive definite. If $A$ commutes with $B^{m}$, prove that $A$ commutes with $B$.

So I don't really know where to start with this other than the decomposition theorem for symmetric positive operators, so that I can write $B$ as diagonal with positive real entries. Perhaps taking the $m^{th}$ root could be useful which we can do since $B$ has positive entries, but I'm not sure.

Many thanks!

TheManWhoNeverSleeps
  • 1,636
  • 3
    It is enough to prove that there exists a polynomial $p \in \mathbb{R}\left[x\right]$ such that $B = p\left(B^m\right)$. To prove this, we can WLOG assume that $B$ is diagonal (since $B$ is diagonalizable). Now it is clear that we can set $p$ to be any polynomial which sends the $m$-th powers of the eigenvalues of $B$ to said eigenvalues. Do you know why such a polynomial exists? (You still need to use positive definiteness -- make sure that you do so, or else the proof is wrong.) – darij grinberg Dec 09 '14 at 04:29
  • I've been trying, but no progress. Would I try to use the minimal polynomial for $B$ or $B^{m}$ somehow? Or should I be trying to just write it out from scratch? – TheManWhoNeverSleeps Dec 09 '14 at 05:01
  • This last step is not really a matrix problem. Do you know about Lagrange interpolation? – darij grinberg Dec 09 '14 at 05:07
  • Reading about it now on wikipedia. So I would use Lagrange interpolation to get a polynomial $p(x)$ such that $p(\lambda^{m})=\lambda$, but how does positivity come into this? – TheManWhoNeverSleeps Dec 09 '14 at 05:14
  • Yes, that's the question :). What is the main condition for Lagrange interpolation to work? – darij grinberg Dec 09 '14 at 05:16
  • does this help ? – r9m Dec 09 '14 at 05:20
  • So, what I've read about the interpolation procedure, nothing is mentioned about when it fails. I would guess some kind of problem occurs when not all of the points have the same sign, but that is as much as I can guess. Granted, perhaps I just skimmed the articles on wiki too much and need to read them in more detail. Edit: Ah, is it that if the points have different sign there are some extra roots involved which are undesirable? – TheManWhoNeverSleeps Dec 09 '14 at 05:35
  • As for the link provided by r9m I will read it in detail tomorrow, thanks :) – TheManWhoNeverSleeps Dec 09 '14 at 05:43
  • No, signs are not the issue per se. But you cannot find an interpolating polynomial if you want two different values at the same point. Why will this not happen here? – darij grinberg Dec 09 '14 at 13:19
  • Well, if say $\lambda$ was negative and $m$ was even then $\lambda^{m}$ would be positive, hence the existence of the polynomial we want would not be guaranteed. – TheManWhoNeverSleeps Dec 10 '14 at 00:29

1 Answers1

3

Since, $B$ is symmetric and positive definite, it is Unitarily diagonalizable, that is $\exists \, U \in U_n(\mathbb{R})$, such that $U^{*}BU = D[d_1,\cdots,d_n]$ is a diagonal matrix, with $d_i > 0$, for all $i = 1(1)n$.

Denote, $U^{*}AU = W = (w_{ij})_{n \times n}$, then,

$AB^m = B^mA \implies U^{*}AU(U^{*}BU)^m = (U^{*}BU)^mU^{*}AU \implies WD^m = D^mW$

So, $d_i^mw_{ij} = w_{ij}d_j^m$ for all $i,j$

Which implies, $d_iw_{ij} = w_{ij}d_j$ for all $i,j$ (since, $d_i,d_j >0$ and $d_i^m = d_j^m \implies d_i = d_j$)

That is $WD = DW$, which in turn implies $AB = BA$.

r9m
  • 17,938